Heading to #SaTML2025 this week to present our work on "Private Selection with Heterogeneous Sensitivities". arxiv.org/abs/2501.05309…

Glad to introduce our new work "Game-Theoretic Regularized Self-Play Alignment of Large Language Models". arxiv.org/abs/2503.00030 🎉 We introduce RSPO, a general, provably convergent framework to bring different regularization strategies into self-play alignment. 🧵👇

CDT student Lorenz Wolf has recently returned from the #SaTML2025 3rd IEEE Conference on Secure and Trustworthy Machine Learning in Copenhagen where he had the opportunity to present his work - more on our blog here ➡️ blogs.ucl.ac.uk/faicdt/2025/04…

🚨 New update from our AI for Cyber Defence for Critical Infrastructure mission The Alan Turing Institute: We're pushing the limits of protocol-aware deception using AI. Here’s what we’ve built👇 airgapped.substack.com/p/update-may-2… Btw this carries minimal dual-use risk. It's a defence-only use case.

RL is the agent-environment loop and we currently do not have enough environments! At Vmax we're building a platform for environment creation.

At RLDM this week to present our work on incorporating diverse prior knowledge in RL (sample efficiency, safety, interpretability,...) Poster #94 on Thursday Full paper here: arxiv.org/abs/2306.01158 #RLDM2025

Had a great time at RLDM this past week! Enjoyed many great conversations and fascinating talks Ben Eysenbach, Michael Dennis, Dan Haramati, DeepFlow. Thanks to The Alan Turing Institute and UKRI CDT in Foundational AI for the support!

🧶1/ Diffusion-based LLMs (dLLMs) are fast & promising—but hard to fine-tune with RL. Why? Because their likelihoods are intractable, making common RL (like GRPO) inefficient & biased. 💡We present a novel method 𝐰𝐝𝟏, that mitigates these headaches. Let’s break it down.👇

*sniff* *pulls shirt* You know, this is perfect - *gestures wildly* - this is the ultimate perversity of capitalism at its purest. Here we have Anthropic, this company claiming to build "AI for humanity," and what do they do? They create this digital cocaine, this Claude Code,

Very excited about our proposal and grateful to be supported by the NVIDIA Academic Grant!!

Post-training methods like RLHF improve LLM quality but often collapse diversity. Check out DQO, a training objective using DPPs that directly optimizes for semantic diversity and quality.

(1/4) Typical LLM post-training mechanisms have a hard time learning models that can produce diverse responses. To fix this we introduce 𝐃𝐐𝐎 (𝐃𝐢𝐯𝐞𝐫𝐬𝐢𝐭𝐲 𝐐𝐮𝐚𝐥𝐢𝐭𝐲 𝐎𝐩𝐭𝐢𝐦𝐢𝐳𝐚𝐭𝐢𝐨𝐧), a method for post-training LLMs to generate diverse high-quality

New pre-print: arxiv.org/abs/2509.26327. The information bottleneck offers useful but imperfect insights into deep learning. Mirco Musolesi, Steve Hailes, and I introduce a Generalized IB that fills the gaps left by the classic approach.

Our seminar is back! The next seminar is Wednesday (Oct 22) and starts at 12:30pm UK time! Arthur Gretton from UCL / Google deepmind is going to talk about “Learning to Act in Noisy Contexts Using Deep Proxy Learning”! ucl.zoom.us/j/99748820264 ucl-ellis.github.io/jt_csml_semina…